The institutional review boards at all participating institutions (Seoul Metropolitan Government Seoul National University Boramae Medical Center, Gangnam Severance Hospital, Seoul National University Bundang Hospital, Catholic University of Korea Incheon St. Mary’s Hospital, Catholic University of Korea Seoul St. Mary’s Hospital, and Korea Cancer Center Hospital) approved this study. Representative institutional review board approval was granted by Seoul Metropolitan Government Seoul National University Boramae Medical Center (H-10-2020-195).

Images were collected from 6 medical institutions in captured DICOM file format (Figure 1). Of the 6 institutions, 3 used iU22 systems (Philips Healthcare), one used EPIQ 5G (Philips Healthcare), one used Prosound Alpha 7 (Hitachi Aloka), and one used Aplio 500 Platinum (Toshiba Medical Systems). Experienced surgeons at each institution labeled the images as benign (fine-needle aspiration cytology Bethesda Category II or benign surgical histology) or malignant (fine-needle aspiration cytology Bethesda Category V/VI or surgical histology of thyroid carcinoma). The images were cropped into 299×299 pixels to include typical thyroid features. The images were not augmented.

Table 1 summarizes details of the thyroid ultrasound images used in this experiment. We used 80% of each institution’s data as training data and the remaining 20% as test data. We used stratified random sampling to select the test data set. There was a total of 4300 malignant images and 2465 benign images in the total training data set and a total of 1075 malignant images and 617 benign images in the test data set. For external validation, 100 thyroid ultrasound images (50 malignant image data and 50 benign) were provided by a medical institution in Japan. We were blinded to the labeling (malignant or benign) of the images.

In addition, to verify the performance of federated learning with external data, we collected an external test data set, which consisted of 50 malignant and 50 benign ultrasound images taken using a TUS-A500 system (Toshiba Medical System) from Kuma Hospital.

We conducted federated learning experiments (Figure 2) with each institution’s serverworker (a computer system that can train deep learning algorithms with local data in the federated learning process) and the coordinator of Seoul National University Hospital to validate federated learning in a real health care environment (serverworker system at each institution: Intel 4-core 2.3 GHz i5-8259U processor, 16 GB DDR4 RAM memory, and 11 GB Nvidia RTX 2080 Ti graphics; coordinator system: 2.3 GHz i5-8259U processor, 16 GB DDR4 RAM, and 8 GB Nvidia GTX 1080). Network training was performed on the serverworkers, and then each serverworker was configured with a high-memory graphic process unit. We configured the system using the processor and external graphics processing unit for system portability. All versions of software (Python version 3.6.5; PyTorch version 1.4.0; PySyft version 0.2.5) were identical between institutions. We installed Ubuntu 18.04 LTS version on each serverworker and the coordinator system.

We used 5 deep neural network classifiers for thyroid ultrasound image analysis: VGG19 [8], ResNet50 [9], ResNext50 [10], SE-ResNet50, and SE-ResNext50 [11]. We also used these 5 models to verify federated learning.

Stochastic optimization (ADAM) was used with the following parameters: β₁=0.9, β₂=0.999, ϵ=10^–8 [12]. The initial learning rate was 0.001 which was reduced by half every 30 rounds. The mini-batch size was 32. We used a binary cross-entropy loss function to train all networks. We trained the network for 120 rounds. We used PyTorch [13] and PySyft [14] to implement and train all networks with federated learning.

After removing all patient identifying information, images from each participating institution were collected at Seoul National University Hospital to create a pooled data set. We used the pooled data set to conduct conventional deep learning. All settings were the same as those for federated learning, with the exception of those used in PySyft, and the same equipment, with the same specifications as those of the serverworker, was used. Only training data from each hospital used in the federated learning were pooled and used for conventional deep learning. The test data set was the same as that used for federated learning.

For the internal test data set, consisting of 1691 images (1075 malignant and 616 benign), and federated learning–trained deep learning algorithms, the accuracies of VGG19, SE-ResNet50, ResNet50, SE-ResNext50, and ResNext50 were 79.5%, 77.9%, 77.4%, 77.2%, and 73.9%, respectively (Table 2; Table S1 in Multimedia Appendix 1). Figure 3 shows the receiver operating characteristic curve [15] of each network for the internal test data set. Area under the receiver operating characteristic (AUROC) curve values of SE-ResNext50, ResNext50, VGG19, SE-ResNet50, and ResNet50 were 87.6%, 86.0%, 82.0%, 79.9%, and 78.9%, respectively.

For the external test data set and federated learning model, the accuracies of ResNet50, SE-ResNet50, VGG19, SE-ResNext50, and ResNext50 were 76.0%, 73.0%, 69.0%, 60.0%, and 56.0%, respectively (Table 3; Table S2 in Multimedia Appendix 1). AUROC curve values of SE-ResNet50, SE-ResNext50, ResNext50, ResNet50, and VGG19 were 86.7%, 83.4%, 83.0%, 81.0%, and 75.2%, respectively.

For each deep learning algorithm trained with the pooled data, the accuracies of VGG19, ResNet50, ResNext50, SE-ResNet50, and SE-ResNext50 were 81.5%, 78.7%, 85.2%, 83.2%, and 85.2%, respectively (Table 4; Table S3 in Multimedia Appendix 1). The AUROC curve values of VGG19, ResNet50, ResNext50, SE-ResNet50, and SE-ResNext50 were 87.6%, 82.6%, 91.0%, 84.5%, and 91.5%, respectively.

For conventional deep learning using the pooled external test data set, the accuracies of VGG19, ResNet50, ResNext50, SE-ResNet50, and SE-ResNext50 were 71.0%, 77.0%, 80.0%, 66.0%, and 76.0%, respectively (Table 5; Table S4 in Multimedia Appendix 1). The AUROC curve values of VGG19, ResNet50, ResNext50, SE-ResNet50, and SE-ResNext50 were 79.3%, 81.2%, 89.7%, 73.4%, and 91.0%, respectively.

The goal of this study was to verify the performance of federated learning in a real-world health care environment. We first collected thyroid nodule data from 6 institutions and designed a federated learning system using these data. We trained each deep learning algorithm (VGG19, ResNet50, ResNext50, SE-ResNet50, and SE-ResNext50) with the federated learning system. We also trained the same deep learning algorithms using conventional deep learning techniques and compared the performance of federated learning with that of conventional deep learning.

The medical vision community is currently actively conducting diagnosis using computer-aided diagnosis [16]. To improve the performance of computer-aided diagnosis, several deep learning algorithms have been developed and applied [17-20]. Various challenges for deep learning with open data sets have been identified [21,22]. In particular, due to health care data privacy regulations, most open data sets only have a small amount of data collected from a single institution. When training and validation are carried out with only a small volume of data, the performance of a deep learning model cannot be properly evaluated, and generality cannot properly be validated. Federated learning, which can train a deep learning model without centralized data, offers a training strategy that addresses these challenges.

There have been several recent reports of the use of federated learning trained with general images [4] and medical imaging [5,6]. McMahan et al [4] published a study using federated learning with federated averaging and reported that the average parameters trained from each serverworker each round performed similarly to those of conventional deep learning and better than those of federated stochastic gradient descent; however, the study used a relatively simple model and general image data sets (MNIST and CIFAR-10). Sheller et al [5] compared federated learning, institutional incremental learning (IIL), and cyclic IIL using the BraTS 2018 data set [21]. IIL is a collaborative learning process that trains a network with data from one institution and then continues training with another institution’s data successively. One disadvantage of this model is that when the network is trained using data from another institution, the patterns trained from the previous institutions’ data are disregarded. To compensate for this shortcoming, Sheller et al [5] proposed cyclic IIL which repeats the IIL process. They used U-Net architecture [17] for brain tumor segmentation with federated learning, IIL, and cyclic IIL and demonstrated that the performance of federated learning was superior to those of IIL and cyclic IIL; however, the study applied federated learning but did not address the class imbalance or data volume imbalance problems associated with federated learning. Li et al [6] also used the BraTS 2018 data set to compare federated learning and centralized data training; they found no significant difference in performance between federated learning and centralized data training. Most federated learning studies compare federated learning with conventional deep learning only, and there are no studies using clinical data from a real-world health care environment.

The application of federated learning in our study shows that this technology has substantial potential applicability in clinical environments. First, federated learning showed performance comparable with that of conventional deep learning, despite an extremely uneven distribution of data volume from each institution. The difference between the hospital with the most data and the hospital with the least data was 17.5 fold. Moreover, the distribution of benign and malignant images was also skewed. For example, the ratio of malignant to benign images was 47:1 for institution 3, whereas it was 1:2 for institution 1. Because data distributions between hospitals are diverse, the conditions presented in this study demonstrated the applicability of federated learning in the real world and its ability to facilitate collaboration between different size institutions.

In medical image analysis, if the amount of data is insufficient, overfitting (learning from noise in data) often occurs. In such cases, only the accuracy of the internal data set is high, and deep learning algorithms cannot be rigorously evaluated. We were able to overcome the issue of overfitting by collecting images from multiple institution and by performing external validation using images from an institute in a different country. We demonstrated that federated learning is able to maximize the efficiency of medical resources and generalizability of deep learning algorithms using data from different size medical institutions (with various imaging devices and different patient groups). This represents scenarios in real-world health care environments [23-26].

In our study, federated learning training took at least 4 times longer than that of conventional deep learning. The training time for federated learning varied depending on the peripheral environment such as internet speed and temperature of graphics process unit. The performance of federated learning may be enhanced with more images or data augmentation. The ideal volume of data and the distribution of data contributed by each institution for peak performance of federated learning is also not yet known. Further investigation into the optimal training environment, training time, data volume, data distribution, and state-of-the-art deep learning algorithms is required for federated learning.

As shown in Table 5, we noted that when thyroid nodules were classified by a conventional deep learning model, the number of malignant calls was extremely high. The same trend is frequently observed in the literature [20,27-29]. As shown in Table 3, we also found this trend to be prominent in federated learning. Because deep learning is a black box [30], we were unable to determine the potential reasons for this tendency, but we plan to investigate this phenomenon in the future.

This study has several limitations. First, we presented the results of federated learning used in a specific context in terms of the number of participating institutions, and the number and ratio of benign and malignant images. Thus, the generalizability of the results in terms of the performance of federated learning is not known and warrants further investigation. We also used thyroid ultrasound images, which are relatively easy to analyze compared to those from computed tomography, magnetic resonance imaging, and histopathology sections. Results may not be generalizable across different imaging modalities. In future work, comparisons of federated learning with unequal data distribution, data augmentation, one-shot learning are required to explore the implications of data imbalance.

We demonstrated that the performance of federated learning using a shared training model and parameters from 6 institutions was comparable with that of conventional deep learning using pooled data. Federated learning is highly generalizable because it can effectively utilize data collected from different environments despite data heterogeneity. Federated learning has the potential to mitigate many systemic privacy risks by sharing only the model and parameters for training without the need to export existing medical data sets.